Research On Labeling Method Of Training Dataset Based On Surveillance Video

With the continuous development of information technology and the explosive growth of information,it is particularly important to deal with the massive amount of data and extract the information that users need.Moving object detection is a research hotspot in the field of computer vision.It is the basis of intelligent video surveillance systems,target tracking,behavior recognition and other technologies.It plays a vital role in the fields of national and social public security,traffic control,and human-computer interaction.However,the amount of surveillance video data collected from the monitoring device is huge and the data redundancy is serious.Users who use surveillance video tend to only care about video segments that have moving objects.Extracting the key video segments containing moving targets is the basis of subsequent video processing;Current moving object detection uses neural network for training,which greatly improves the accuracy of target detection.However,as the detection model expands,the required training dataset also grows rapidly.Repeated manual labeling training dataset is high and inefficient.In view of the above problems,our paper designs a semi-automatic labeling method based on monitoring video for training dataset,which is mainly divided into the following contents:Firstly,for the key video segment extraction of surveillance video,our paper proposes a moving object detection algorithm based on edge feature and multi-frame differential mixed Gaussian model.In view of the fact that the interframe difference method is not sensitive to noise such as sudden changes in light,our paper integrates it into the Gaussian background model update,divides the video frame into the background area and the foreground area,reduces the Gaussian model update rate of the background area,and improves the motion area.The Gaussian model update rate is dynamically adjusted according to the pixel division.For edge extraction,the eight-direction translation method is used in this paper.This method improves the shortcomings of the traditional edge detection target with unclear outline and poor sealing.Finally,the morphological processing of background noise and foreground holes is used.The foreground target extracted by the inter-frame difference method,the result obtained by the Gaussian mixture model,and the edge information of the object obtained by the translation method are phase-ANDed to obtain the final detection result.Secondly,based on the extracted key video segments,a semi-automatic labeling method based on video object segmentation technology is designed.Visual saliency or motion saliency methods cannot provide complete segmentation.In our paper,as an initial foreground possibility,SLIC superpixel segmentation is used to segment the image into compact regions that do not overlap each other.Each superpixel is mapped to the feature space using an RGB color histogram,an LAB color histogram,and an HOG feature descriptor.The N nearest neighbor regions of each superpixel region are found,and the foreground possibility of each region is updated by the weighted average of the foreground possibilities of the N nearest neighbor regions,and the segmentation result of the final whole video is obtained.Finally,the experimental environment was set up,and the improved moving object detection,video object segmentation method and semi-automatic labeling method were respectively tested.Analysis was performed on accuracy,recall,and F value.